AI does the research and writes the draft. It does not decide what gets published.
Episodes are researched from cited sources and, by default, the whole script is checked against them; every one is cleared by a person before it is narrated and again before it is released. Here is the whole process.
- Researched from real sources, cited on the episode page where you can read them.
- By default, the whole script is checked against those sources before anything is recorded.
- A person clears every episode twice — once before narration, once before release.
- Nearly all of it leaves evidence on the episode page, so you can check rather than trust.
The panel at the top of every episode lists its sources and its transcript.
Most people hear “AI-generated podcast” and picture a model asked for ten minutes on a topic, published unread. That is a fair thing to assume, and it is not what happens here. Every step below is something the pipeline actually does, and most of them leave evidence on the episode page itself — so you can stop reading this and go and check.
What happens, in order
It starts with sources, not with a prompt
Before anything is written, the topic is researched. Searches are run, and every result that is kept is fetched — read in full wherever the page allows it, rather than summarised from a search snippet. The whole page is kept, not just the first screen of it, up to a length that holds a typical journal article, so that a paper’s results table is still there to be checked against later and not just its abstract. A document longer than that is kept only up to that point, and whoever reviews the episode is told when a check relied on one. Whole classes of low-evidence source are kept out of the search itself: discussion forums, fan and how-to wikis, video platforms, and retailers selling the thing under discussion. That list applies to searches as they are run, so an episode published before a domain was added to it can still cite one. A client can supply their own material instead, a report or a transcript or a set of articles, and then that is the research.
Either way, the sources end up listed on the finished episode’s public page, linked wherever there is a page to link to. Not a count, and not a claim that research was done — the sources themselves, for anyone who wants to go and read them. A source that was only ever a search result stays on that list too, rather than being quietly dropped to make the list look uniformly thorough.
The script is written from that research, under its category’s rules
The draft is written from the material that was gathered, not from a model’s recollection of the subject. If the episode falls in a regulated category, the rules for that category are part of the instructions before the first word exists: a health episode is told to report what a trial used rather than what a listener should take, and to close on what is still genuinely unsettled; a money episode is told not to recommend buying or selling anything, and to present figures as observation rather than prediction.
Those constraints shape the draft. They are not a warning label bolted onto the end of one.
By default, the finished script is checked against the captured source text
When the draft is finished, every sentence of the script, the hook and the show notes is checked, by default, against the source text captured for that episode — by a separate AI model, never the one that wrote it. Sentences a source contradicts are highlighted, and so are high-stakes claims — a dose, a safety risk, a money or legal rule — that nothing in the sources supports. A quote the check relies on is confirmed by code to really be in the source. This is a check against the research the episode was actually built on, not a second opinion on whether a sentence sounds plausible, and it can miss things: how often, in our own test, is on the fact-check page.
A highlight is not a note in a report that someone might read. Until every one is cleared — the sentence accepted, rewritten, or cut — the episode cannot be narrated at all; the request is refused, and so it is if the check itself did not finish. Editing the script afterwards runs the check again, so a fix cannot quietly introduce a new unsupported line on its way past the gate.
Narration and release are two separate decisions, both made by a person
A finished draft goes into review and stays there. Nothing narrates itself out of that state: a person reads the script, edits it if it needs editing, and starts the narration. Publishing is a second decision made later, after the audio exists — an episode that has been fully narrated is still not public until someone releases it, or sets the time it should go out.
That is two deliberate acts between a finished draft and a listener, placed at the two points where a mistake would otherwise reach an audience.
Regulated topics carry their disclaimer in the audio, not in the small print
When an episode’s subject calls for one — health, finance, legal — the disclaimer is read aloud in the episode itself, in the same voice as the rest of it, and printed on the episode page. For health and finance it is switched on the moment the episode is created, before anyone chooses anything; the legal one is put in front of the person reviewing the script, for them to switch on.
That placement is the point. A disclaimer in a website footer is invisible to someone listening in a podcast app, which is where almost everyone listens.
The AI part is disclosed on the episode, not buried on this page
Every episode page states plainly that the episode was researched and scripted with AI assistance and reviewed by a person, and that the narration voice is synthetic. Both statements are repeated inside the feed itself, so they travel with the episode into whatever app someone subscribes in.
Nobody should have to find their way to a methodology page to learn that.
What ships with a published episode
Everything the pipeline produced for an episode ships on its public page: the audio, the show notes, the full transcript and the linked list of sources — plus an entry in an RSS feed that podcast apps, Spotify among them, subscribe to.
The transcript and the sources are on the page for the same reason the disclaimer is in the audio: evidence is only worth anything where the listener already is.
Go and check one
All of this is just a claim until you open an episode and look. The panel at the top of the episode page lists the sources it cites and the transcript it ships with; the disclaimer, where the subject needs one, is on the page and spoken before the main discussion; and the sources, where there is a page to link to, are linked for anyone who wants to read them.
Open a published episode: Berberine, Blood Sugar, and the "Nature's Ozempic" Problem